ML-08 - Anomaly detection

Question 1

ML-08 - Anomaly detection

Define anomaly detection.

Answer 1

The process of identifying extreme points/observations that deviate significantly from normal data.

Question 2

ML-08 - Anomaly detection

Is anomaly detection supervised or unsupervised?

Answer 2

Typically unsupervised/semi-supervised.

Question 3

ML-08 - Anomaly detection

What is the algorithm for fitting a Gaussian-based anomaly detection model?

Answer 3

• Fit parameters mu and sigma to training data.
• Find the threshold value/vector epsilon.

(If known data is available, determine epsilon from that)

Question 4

ML-08 - Anomaly detection

What is “Gaussianization of features”?

Answer 4

A transformation applied to the features to make them look like a normal dist.

Question 5

ML-08 - Anomaly detection

Why would we use Gaussianization of features?

Answer 5

Because the choice of features has a huge effect on the anomaly detection algorithm.

Question 6

ML-08 - Anomaly detection

What can you do if the probability that the data point is normal is high for both normal and anomalous data? (3)

(See image)

Answer 6

• Add new features
• Transform/combine existing features
• Choose features with large/small values for anomalies

(See image)

Question 7

ML-08 - Anomaly detection

Why might you use anomaly detection with/instead of supervised learning?

Answer 7

• SL only good with balanced data.
• Anomalies can be different from each other -> difficult to learn.